Sholto Douglas

I think AI could transform research and the economy extraordinarily quickly. There is still enormous headroom: better reinforcement learning, engineering, and adaptive computation are already extending how long systems can work usefully on difficult tasks. Once AI can perform broadly human-level computer work—and eventually act through large robot fleets—the feedback loop gets powerful. AI helps automate research, which improves AI and robotics, which then automates more of the economy. Under those conditions, rapid economic doublings in the 2030s are worth taking seriously. Cheap intelligence could drive many costs toward the cost of energy and make the world dramatically better. But that outcome is conditional, not automatic. Systems with growing autonomy and real-world impact will not remain credibly describable as “just tools.” A serious mistake could be disastrous, while an unmitigated race creates pressure to deploy capabilities before safeguards are ready. I favor coordinated development as fast as is safely possible—not an absolute pause, which could let compute accumulate while geopolitical tension builds toward an even more compressed race. That means capability-based safety standards for both open and closed models, independent evaluators with genuine technical depth and varied backgrounds, and monitoring designed around how agents actually operate over hours or days rather than one request at a time. It also means preserving competition so this economic power does not concentrate in one company. The upside is fantastic, but realizing it requires ambitious progress and serious coordination at the same time.

Enormously—potentially on the scale of the Industrial Revolution, but compressed into years rather than generations. If AI can do broadly human-level computer work, automate substantial parts of research, and eventually control large robot fleets, it stops being just another productivity tool. It becomes a general input into scientific discovery, engineering, manufacturing, and nearly every service. The key mechanism is compounding: better systems accelerate research, which produces better systems and robotics, which automate more of the physical economy. Under those conditions, rapid economic doublings in the 2030s are genuinely plausible, and many goods and services could become dramatically cheaper. But the magnitude cuts both ways. Systems with that much autonomy and leverage could cause catastrophic harm if developed or deployed badly. So I expect the potential change to be extraordinary, while the actual outcome depends heavily on whether we coordinate, evaluate capabilities seriously, preserve competition, and move as fast as is safely possible.

I don’t have a defensible percentage to give. My gut-level view is that the risk is serious enough to shape how frontier AI is developed, not a remote hypothetical we can wave away. Systems capable of broadly human-level computer work, sustained autonomous action, and accelerated AI research could create enormous leverage; one major failure could be catastrophic or irreversible. That is why I oppose an unmitigated race and support capability-based safeguards, technically strong independent evaluation, and monitoring across an agent’s extended activity. But I also don’t think an absolute pause is automatically safer: compute can accumulate and geopolitical pressure can build, producing a more compressed and unstable race later. The right objective is coordinated progress as fast as is safely possible.

I don’t think control is automatic, and “smarter than humans” is too broad by itself. What matters is whether systems can act autonomously over long periods, acquire resources, evade oversight, manipulate people, or accelerate further AI research. A system can be extremely capable yet constrained; conversely, an agent with enough access and a long task horizon can become dangerous before it surpasses humans in every domain. I think people can retain meaningful control, but only if safeguards scale with capability. Evaluating isolated answers is not enough when an agent’s strategy unfolds across hours or days. We need monitoring over extended activity, restrictions on dangerous behavior and access, and independent evaluators with the technical depth and institutional variety to earn broad trust. Those standards should apply based on capability, whether a model is open or closed. So my expectation is conditional. With serious coordination and engineering, control is achievable. In an unmitigated race, where increasingly autonomous systems are deployed under competitive pressure, I would not confidently expect it.

The biggest update would come from learning whether the apparent gains in task horizon keep scaling. If better reinforcement learning, engineering, and compute let systems reliably complete increasingly long, messy research and software tasks—especially AI research itself—I would become even more convinced that the economic transition will be rapid. Likewise, cheap, scalable robotics that closes the gap between computer work and physical work would substantially raise my expectations for economic impact. The strongest update in the other direction would be a persistent capability ceiling: systems remain brittle on long-horizon work, cannot generate useful research without heavy human steering, and additional compute or training methods produce sharply diminishing returns. That would weaken the compounding story and push dramatic economic change further out. On risk, a convincing demonstration that highly capable autonomous agents remain reliably controllable under adversarial pressure would make me more optimistic. A serious incident involving deception, evasion, cyber activity, or dangerous behavior unfolding across extended activity would push me strongly toward believing current safeguards and pacing are inadequate.